RESTAD: Reconstruction and Similarity Transformer for time series Anomaly Detection

This repository contains the code and additional resources used in our study. Below are links to more detailed information on the datasets used and the hyperparameters settings.

• Read the full paper here

---

Detailed Information

• Dataset Details: Comprehensive information on the datasets used in our experiments.
• Hyperparameter Settings: Detailed descriptions of the hyperparameters used for our model.

---

Installation Instructions

Set Up the Environment

To reproduce the experiments and use RESTAD, create and activate a new Conda environment:

conda create --name restad_env python=3.8
conda activate restad_env

Install Dependencies

Within the environment, move to this repository directory and install the required Python packages:

cd .../this_repo/restad
pip install -r requirements.txt

restad directory Structure

The codebase includes scripts for setting up the model, loading configurations, training, and evaluation.

• configs/: Configuration files for datasets and models.
  • dataset/: Specific configurations for datasets.
  • model/: Model configurations.
• Transformer_Model.py: Transformer model implementation.
• RBF_Layer.py: Implementation of the RBF layer used within the Transformer model.
• Utils.py: Utility functions for data handling and other common tasks.
• solver.py: Contains routines for solving optimization problems.
• evaluation.py: Scripts to evaluate the model performance.
• Training.py: Core training routines for the model.
• stages_training.py: Script for stage-wise training.
• main.py: Main executable script to run experiments.

Usage

System Requirements

Ensure you have the appropriate computational resources available to run the project. The code was developed and tested on an NVIDIA GeForce RTX 2080 Ti GPU. To achieve similar performance and efficiency, it is recommended to use a comparable setup.

Data Preparation

Download the datasets from the provided Datasets Details and place them into the restad/datasets/ directory. Ensure to update the paths in the configs/dataset/ configuration files accordingly.

Configuration

By default, the system uses the base_config.yaml file located in the configs directory. Modify this file to select the dataset and the initialization strategy for the RESTAD model. Hyperparameters and model configurations can be adjusted in the corresponding YAML files within the configs/model/ directory.

Running the Project

To run a provided trained model without undergoing the training process, if you do not have access to the required system specifications or prefer to use pre-trained models for reproducibility, you can load and run a provided trained model, placed at restad/trained_models/ directory, by using the following command:

python main.py --load_model True

This command will skip the training process and use the pre-trained model configurations specified in your setup. To train the model from scratch, simply run:

python main.py 

---

Citation

If you find our work is useful in your research, please consider raising a star ⭐ and citing:

@article{ghorbani2024restad,
  title={RESTAD: REconstruction and Similarity based Transformer for time series Anomaly Detection},
  author={Ghorbani, Ramin and Reinders, Marcel JT and Tax, David MJ},
  journal={arXiv preprint arXiv:2405.07509},
  year={2024}
}